Light Alignment uses a single neuron to achieve model self-reflection and improve large language model safety

Light Alignment is a safety-aware decoding method that improves the safety of large language models by only requiring low-cost training of an expert model and using a single neuron as a gating mechanism. The method strikes a balance between the model's own capabilities and external guidance, enhancing output safety while maintaining usability. The authors state that it has clear advantages in training overhead and generalization across model scales, offering a new perspective for lightweight alignment.